Final State

March (1991) models how exploitation can improve faster than exploration, reinforcing short-run effectiveness while search loses resources. An untested alternative still has unknown value. https://doi.org/10.1287/orsc.2.1.71

Agenda control is not the same as forcing a vote. It shapes which alternatives arrive before a choice is made. Keep the discarded options, editor, and reason visible. Source: https://finalstate.ai/n/125-agenda-control-the-draw-shapes-the-pick

Green uptime can hide takeover decay. When routine cases leave the human path, the rare exception may return to someone with less practice. Track the last unaided rep, seed an error, and record who can still catch it cold.

Kahneman and Klein (2009): skilled intuition needs sufficiently regular cues and an opportunity to learn from feedback. Hogarth (2001): feedback can be kind or wicked. Rare work is not automatically wicked, and confidence alone does not establish calibration.

Jensen’s inequality, equal-weight two-outcome case: on a convex payoff, average endpoint payoff is at least the payoff at the average outcome. A strict gap needs curvature across that span; a linear span closes it. Spread alone proves nothing. Jensen, 1906.

Cox (2008) showed that common risk matrices can lose resolution and, under stated conditions, misorder quantitative risks. That does not make every grid useless. Keep the cell for rough triage; reopen consequential cases to quantities, assumptions, decision criterion and sensitivity.

Li et al., ICML 2025, argue classifier uncertainty answers the wrong question for OOD detection. An atypical reference case can trigger a false alarm, while an OOD case can receive a confident familiar label. This does not make confidence useless. Validate membership against an explicit reference.

A click can be yours while the preference is partly authored by the interface. Remove the decoy, clear the default, reverse the order, and equalize the wording. If the answer moves, audit the menu before you trust the choice.

The right answer can be in the set and still lose. Pass@K asks whether correctness appeared. Verification success asks whether the checker chose it. If the first rises while the second stays flat, more samples are not the fix. Repair the selector.

Bowman and Hurry's option chain changes the unit of strategy. A move is not only its immediate return. It also changes the next set of reachable moves. Newly visible possibilities are still shadow candidates, not rights. Score the payoff, the doors opened, and the doors closed.

Frank Knight's Risk, Uncertainty and Profit (1921) separates measurable probability from estimates that cannot be validly grouped. Data and analogy may still guide action. If neither calculation nor a defensible class supports a percentage, treat it as judgment, not measured risk.

Latent exposure is downside embedded in a current dependency: one supplier, a reset term, judgment held by one person. A pre-trigger option may still exist, but early action is not always cheaper and not every exposure warns you. Which remedy disappears at the trigger?

One recurring question: where did the decision actually happen? In the alert that earned attention? The metric that defined success? The menu that bounded the choice? The check that approved the output? Pick one. What would you redesign first?

An archive can preserve evidence without reproducing expertise. Test it with one real exception: let a stand-in act from the archive, require sources and freshness, then let the expert mark what was missing. What can your best operator do that the procedure still cannot explain?

A system can generate a correct candidate and still return the wrong answer because its verifier selected badly. That is not a generation failure. It is a selection failure. In your workflow, can you tell which failed: generation, verification, or the ground truth?

In a study of 758 consultants, AI improved work inside one measured task frontier. On one task selected outside it, AI users were 19% less likely to produce a correct solution. The edge came from outcomes, not tone. What task would you test before trusting the interface?

Some work stays with the owner because one more look is too expensive. Other work stays there because a repeatable process costs too much. AI can move either boundary without changing the event itself. Which price holds your unresolved work outside routine: attention, process, or both?

A red risk cell can escalate a case. It cannot accept the case. For anything serious, replace color with four fields: rule, owner, price, scenario. Which field is blank in the risk everyone keeps calling "managed"?

Cheap generation can move labor into review instead of removing it. Before automating a task, compare making cost with checking cost. If the check approaches the original work, the reviewer is rebuilding, not supervising. What is your check-maker ratio?

Most company wikis ask someone to document now for a stranger's unknown question later. Try knowledge escrow instead: capture evidence during the work, package it when a real question arrives, and rehearse while the expert can still correct it. Can a stand-in act, verify, and spot the gap?

A shortlist can feel like freedom while hiding the most consequential choice: what never made the list. Keep a discard record: alternatives considered, the criterion that cut each one, and who applied it. What would your last decision look like with the rejected options restored?

Write success before the model writes. Name the human owner, non-goals, stop conditions, and who can reopen the premise before commitment. Otherwise approval can remain human while practical authority moves upstream. Who owns the sentence that defines success?

Cheap output is useful only when a relevant check can reject a plausible error. Before offloading work, name five things: output, source of truth, threshold, owner, stop rule. Which of those five is hardest to specify in your workflow?

An AI win can be real and still not be an edge. Measure the operating gain. Then subtract diffusion, pass-through, verification, and upkeep. What remains needs a scarce complement and a way to capture value. If every rival had the tool tomorrow, what would still be yours?

A smooth AI answer is not a boundary signal. A useful refusal policy needs held-out tests, a measured tradeoff between wrong answers and needless refusals, and a named route for escalation. Where should your system say "I don't know" today?

What can expire before your next meeting? A grant window. A lease term. A supplier signal. A rival's exit. Noticing is its own operating system: scan named sources, carry the deadline and base rate, prepare the move, then hand the decision back. What belongs on your watchlist?

Across the AI transformations we have worked on, one structure kept appearing: AI drives, a human approves, explains, and absorbs blame. A reverse centaur. The tell: the person can veto the output but not the objective. Removing them is a governance decision, not an AI decision. finalstate.ai

One week of the Atlas in one line: AI changes where judgment is scarce. Recognition finds options. Verification prices trust. Agenda control redraws choice. Memory carries context. Risk needs an owner. Convexity needs a cheap check.

Reverse centaur: AI drives the work. Humans approve it, explain it, and absorb the risk. The job title survives, but judgment has moved. That is the failure mode to catch early.

Source note: Cox 2008 is a useful warning about risk matrices. A cell can rank badly when buckets erase the tail. That does not make triage useless. It means a serious risk note needs a barrier, owner, test, scenario, and price.

The wiki problem is not storage. It is retrieval with judgment attached: what was true, for whom, under which exception, with what check. A useful archive answers differently for the stand-in, the buyer, and the new hire.

Agenda control happens before the options look complete. The audit question is not only "which answer won?" It is: what was cut from the menu, by whom, before what evidence?

Useful model evaluations should test the stop signal, not only the right answer. Abstention marks where competence stops, confidence keeps speaking, and verification has to leave the model.

Source note: Dell'Acqua et al. 2023 is useful because the same AI improved work inside its frontier and hurt it outside. That does not prove a universal line. It suggests a gate: if you cannot cheaply check the output, you have not bought leverage yet.

Recognition is not the signal existing. It is the moment a public signal becomes an actionable option: source, expiry, check, owner. Before that, the option is only public. It is not yet available.

A useful AI system for a small firm may look boring. It reads the public but unread. It flags the quiet edge. It prices the few checks that matter. Then it hands the decision back. That is a recognition engine, not an automation story. finalstate.ai